Text this: Extrapolating training of neural networks